CHECKPOINT.md

Project

Artificial Consideration (formerly 'Artificial Memory')

Core Insight

A second attention process may be more valuable than a larger context window.

The architecture does not attempt to emulate biological memory. Instead, it attempts to emulate consideration:

Given the current topic, what additional information should be present in awareness right now?

---

Architecture

LLM1

Primary generative model.

Responsibilities:

• Conversation
• Reasoning
• Coding
• Planning
• User interaction

Attention represents:

• Current topic
• Current intent
• Current constraints
• Operator goals

LLM2

Artificial Consideration Engine.

Responsibilities:

• Observe LLM1 topics
• Maintain an active subset from a datastore
• Predict likely conversational trajectories
• Stream high-worth context tokens
• Prune low-value context

LLM2 is not primarily a generator. LLM2 is a router.

---

Fundamental Shift

Not vertical scaling:

• Bigger models
• More parameters

Not horizontal scaling:

• More agents
• More model copies

Orthogonal scaling:

• Parallel attentions
• Separate roles
• Symbiotic operation

LLM1 attends to the conversation. LLM2 attends to the datastore.

---

Artificial Consideration

The term 'Artificial Memory' is likely misleading.

Memory implies:

• Storage
• Learning
• Forgetting
• Consolidation

The proposed architecture performs none of these.

Instead:

• Datastore holds information.
• Operator edits datastore.
• LLM2 determines relevance.
• LLM1 receives consideration.

Therefore:

Artificial Consideration or Artificial Influence

are more accurate descriptions.

---

Datastore Principles

Datastore is external.

Models remain stateless.

Datastore changes only through:

• Additions by operator
• Deletions by operator
• Edits by operator

No autonomous mutation. No autonomous learning. No autonomous memory creation.

This simplifies:

• Auditability
• Debugging
• Determinism
• Reproducibility

---

High-Worth Token Injection

The primary mechanism.

Observation:

Context volume is not equivalent to context quality.

Hypothesis:

A small number of extremely relevant tokens may outperform a very large amount of generic context.

Candidate categories:

• Constraints
• Invariants
• Prior failures
• Architecture decisions
• Dependency warnings
• Capabilities
• Safety rules

Injected tokens should be:

• Small
• Relevant
• Actionable
• Timely

---

Continuous Streaming Model

Important refinement:

LLM2 does not query. LLM2 streams.

LLM2 continuously:

• Watches topics
• Monitors shifts
• Swaps active subsets
• Injects considerations

This resembles a perpetual routing process.

---

Trajectory Concept

Possibly one of the strongest ideas discussed.

Rather than only observing current topics:

LLM2 may predict:

'Where does this kind of conversation usually go next?'

Potential applications:

• Coding
• Refactoring
• Architecture design
• Incident response
• Documentation

This creates an attention->trajectory->consideration loop.

---

Strongest Claim

The strongest testable claim is NOT:

'Artificial memory works.'

The strongest testable claim is:

'A second datastore-backed attention process can improve inference quality by injecting a tiny set of high-worth context tokens.'

---

Minimal Prototype

Do NOT build the full architecture first.

Build:

Prompt -> Keyword extraction -> Memory lookup -> Top-N considerations -> Injection -> LLM

If this demonstrates measurable improvement:

• Then introduce LLM2.
• Then introduce streaming.
• Then introduce trajectory prediction.

---

Shelf of Wonder Test

This idea possesses a valuable property:

It is removable.

If useful:

• Keep it.

If not useful:

• Unbolt LLM2.
• Return to baseline architecture.

No retraining required. No damage to core systems.

---

Current Assessment

Status: Worth Prototyping.

Reasons:

• Uses existing hardware.
• Uses existing models.
• Requires minimal architectural intrusion.
• Tests a genuinely different scaling direction.
• Creates a second attention field rather than a larger context field.

Key open question:

Can a tiny stream of carefully selected consideration tokens consistently improve inference quality?

Everything depends on that experiment.